Arbitrarily Small Execution-Time Certificate: What was Missed in Analog Optimization

TL;DR

This paper proposes a novel analog optimization paradigm that leverages differential equations and fixed-time stability to achieve ultra-low energy consumption, scalability, and arbitrarily small execution times with certificates of convergence.

Contribution

It introduces a new analog optimization framework with a fixed-time-stable scheme and infeasibility detection for convex NLPs, enabling certified, scalable, and ultra-fast solutions.

Findings

Transforming optimization problems into ODEs enables real-time analog solutions.

The fixed-time-stable scheme guarantees arbitrarily small execution times.

Infeasibility detection is achieved through a homogeneous monotone complementarity formulation.

Abstract

Numerical optimization (solving optimization problems using digital computers) currently dominates but has three major drawbacks: high energy consumption, poor scalability, and lack of an execution time certificate. To address these challenges, this article explores the recent resurgence of analog computers, proposing a novel paradigm of arbitrarily small execution-time-certified analog optimization (solving optimization problems via analog computers). To achieve ultra-low energy consumption, this paradigm transforms optimization problems into ordinary differential equations (ODEs) and leverages the ability of analog computers to naturally solve ODEs (no need for time discretization) in physically real time. However, this transformation can fail if the optimization problem, such as the general convex nonlinear programs (NLPs) considered in this article, has no feasible solution. To avoid transformation failure and enable infeasibility detection, we introduce the homogeneous monotone complementarity problem formulation for convex NLPs. To achieve scalability and an execution time certificate, this paper introduces the Newton-based fixed-time-stable scheme for the transformed ODE, whose settling time $T_p$ can be prescribed by choosing the ODE's time coefficient as $k=\frac{\pi}{2T_p}$. This equation certifies that the settling time (execution time) is independent of the dimension of the optimization problems and can be arbitrarily small if the analog computer allows.